VoiceDetect AMD: AI-Powered Answering Machine Detection for Call Centers

If you run a call center with VICIdial or any Asterisk-based dialer, you know the pain of Asterisk’s built-in AMD. It’s slow, inaccurate, and hasn’t evolved in years. Agents waste time on voicemails, and good leads get hung up on.

I built VoiceDetect AMD to solve this problem using modern AI and speech-to-text technology.

The Problem with Traditional AMD

Asterisk’s built-in AMD uses audio pattern detection - analyzing silence, tone patterns, and timing. The problems:

• Slow: Takes 4-5+ seconds to make a decision
• Inaccurate: High false positive rates, especially with varied greetings
• No learning: Can’t improve based on your specific call patterns
• Black box: No visibility into why decisions are made

How VoiceDetect Works Differently

Instead of guessing based on audio patterns, VoiceDetect actually listens to what’s being said:

1. Stream audio via WebSocket when call connects
2. Transcribe speech using cloud STT (OpenAI Whisper, Deepgram)
3. Classify transcript with a trained ML model
4. Return decision - Human or Machine - in ~3 seconds

The key insight: “Hi, you’ve reached John’s voicemail…” is obviously a machine. A human would say “Hello?” or “This is John”. By understanding the words, not just the audio patterns, we get much better accuracy.

The Dashboard

Monitor your AMD performance in real-time:

Call Logs - Full Visibility

Every call is logged with complete details:

Search & Filter:

• Filter by Called Number, Caller ID, Lead ID, or Unique ID
• Filter by decision type (Human/Machine)
• Date range filtering for historical analysis
• Export to CSV for reporting

For Each Call:

• Full transcript of detected speech
• Confidence percentage
• Processing latency
• VICIdial integration data (Lead ID, CID)

This visibility is crucial for debugging and improving your classifier.

Train Your Own Classifier

This is where VoiceDetect really shines. Each tenant gets their own ML classifier that you can train:

How Training Works:

1. Upload audio samples or text transcripts
2. Label each sample as Human or Machine
3. Click train - model rebuilds with your data
4. Deploy the new model version

Your classifier learns from YOUR call patterns. Spanish call center? Train on Spanish samples. B2B calls with professional greetings? Train on those. The model adapts to your specific use case.

Test Before You Deploy

Before going live with a new model, test it:

Two Testing Methods:

• Text Input: Paste a transcript to see how it would be classified
• Voice Upload: Upload an audio file for full end-to-end testing

See the classification result and confidence score. Debug edge cases before they affect your campaigns.

One-Command Installation

Deploy to your VICIdial/ViciBox servers with a single command:

1

curl -sSL "https://..." | sudo bash

What Gets Installed:

• Python websocket-client library
• Configuration file at /etc/voicedetect.conf
• AGI script at /var/lib/asterisk/agi-bin/
• Dialplan extension 8370 for AI AMD
• Heartbeat monitoring (every minute)
• voicedetect-test command for verification

Server Monitoring:

• Track all installed servers in the dashboard
• Online/offline status with heartbeat
• View server details (IP, Asterisk version)
• Last seen timestamp for troubleshooting

Fine-Tune AMD Behavior

Customize timing and thresholds to match your needs:

Timing Settings:

• Silence Timeout: How long to wait with no speech (default 2s)
• Speech Processing: Time after speech starts (default 3s)
• Max Total: Hard limit regardless of speech (default 6s)

Decision Thresholds:

• Min Confidence: Only accept decisions above this threshold (default 60%)
• Silence Handling: Treat no speech as Human or Machine

Expected Decision Times:

• Dead air (no speech): ~2 seconds
• Immediate talker: ~3.5 seconds
• Delayed talker: up to 5 seconds
• Hard limit: 6 seconds max

Key Features

No GPU Required

VoiceDetect uses cloud STT APIs (OpenAI Whisper, Deepgram) for transcription. Run on any server without expensive GPU hardware.

Multi-Tenant Architecture

Each customer gets:

• Their own trained classifier
• Isolated call logs and data
• Custom thresholds and settings
• Separate usage tracking

VICIdial Integration

Drop-in replacement for Asterisk AMD:

• EAGI integration
• Works with existing dialplan
• Lead ID and CID passthrough
• Compatible with ViciBox

Real-Time WebSocket Streaming

Low-latency audio streaming means faster decisions. No waiting for the call to buffer - analysis happens as speech is detected.

Technical Flow

1. Call Connects → Audio streams to VoiceDetect via WebSocket
2. VAD Detection → Voice Activity Detection identifies speech
3. STT Transcription → Cloud API converts speech to text
4. ML Classification → Your tenant-specific model analyzes the transcript
5. Decision Returned → Human or Machine result via WebSocket
6. Asterisk Routes → Route to agent or hangup

Total time: approximately 3 seconds from answer to decision.

Use Cases

Outbound Call Centers

Reduce agent idle time by filtering voicemails before connection. More conversations per hour, better agent utilization.

Predictive Dialers

Make smarter routing decisions. Connect humans to agents, drop machines automatically.

Lead Verification

Verify that leads have valid, human-answered phone numbers before adding to campaigns.

Compliance

Some regulations require human confirmation. VoiceDetect provides auditable logs of every classification decision.

Getting Started

VoiceDetect AMD is available as a hosted service or self-hosted solution.

Interested? Get in touch to discuss your requirements and see a demo.

VoiceDetect AMD is part of my VoIP DevOps product suite, alongside eFax for professional fax management.